OpenAI plans to buy Google TPUs as Oracle turns a rival's chip into ordinary cloud infrastructure

OpenAI plans to purchase Google’s tensor processing units, The Information reports, although financial terms, scale and intended workloads are not known. OpenAI declined to comment and Google did not confirm the prospective transaction. The evidence around the direction is stronger than the missing deal details: OpenAI is hiring an engineer to optimize its frontier models for TPUs and already rents some Google computing capacity.

The more structural development is public. Google and Oracle announced that Oracle Cloud Infrastructure will begin offering TPUs in 2027, making OCI the first hyperscale cloud outside Google to carry them. Oracle will provide the data centres, cooling, networking and proximity to its enterprise databases; Google supplies a chip-and-software system that it previously kept primarily inside its own cloud.

That changes the competitive unit. Nvidia’s advantage is not only processor performance but an ecosystem that lets customers acquire GPUs from almost any major cloud and use familiar software. A Google accelerator confined to Google Cloud asks buyers to accept both a different architecture and a rival platform. Putting TPUs inside Oracle separates those choices. OpenAI can diversify its compute without locating every workload in Google’s cloud, while Oracle can sell a second architecture alongside Nvidia GPUs.

This is not an Nvidia exit. OpenAI has enormous Nvidia-linked commitments, including an Ohio project supported by Nvidia credit, and its own custom chip program with Broadcom. Porting frontier training and inference is also harder than moving a generic application: compiler behavior, kernels, networking and model design all affect utilization. The new job posting suggests that optimization work is still early.

The strategic value lies in credible alternatives. If OpenAI can run important workloads across Nvidia GPUs, Google TPUs and its own silicon, no single supplier defines its cost curve or delivery schedule. For Google, selling chips to the company challenging Gemini turns a product rivalry into infrastructure revenue. Oracle is the intermediary that makes that contradiction commercially workable.